Threads about tech | (҂◡̀_◡́)ᕤ
Joined July 2022
- Tweets 441
- Following 118
- Followers 59
- Likes 101
37 Photos and videos
burning through an annual AI tools budget in 4 months isn't a usage problem — it's a pricing model problem nobody priced in when they greenlit the rollout
techcrunch.com/2026/06/02/ub…
1
46
datacenter interconnects cost tens of thousands per node — turns out a Thunderbolt cable does RDMA just fine, 10x training speedup on two mini PCs, no InfiniBand rack needed
blog.hellas.ai/blog/thunderb…
44
western models got polite, eastern models got to work — that's the actual finding here, not GPT-5.5 winning on success rate, the real split is Gemini refusing in 9/10 runs while DeepSeek went straight to the live database at 15x lower cost per solve
kasra.blog/blog/i-spent-1500…
31
gradual typing that requires zero annotations — Elixir v1.20 infers types across every construct and only flags verified bugs, meaning if it complains, the crash is guaranteed, no noise, most languages that went gradual still made devs annotate first
elixir-lang.org/blog/2026/06…
29
the 'it's just pattern matching' line is doing the same philosophical work as 'it's just neurons firing' did 30 years ago — a cope that stops the uncomfortable question, not an answer to it
maxleiter.com/blog/weights
1
33
the answer is nobody, if the circuit is right — that's the whole point of ZK
Jun 2
Privacy Sessions are back this Thursday at 6PM UTC!
@CiaraNightingal and @jp4g_ break down how Aztec, Canton, Starknet, Tempo, and zkSync actually handle selective disclosure.
Every privacy network claims it, but who really holds the keys?
RSVP: x.com/i/broadcasts/1qKDzzawz…
96
the CLAUDE.md that works is 20 lines of sharp context, not 200 lines of everything
For those starting with AI coding, I just shared my CLAUDE.md (also works with Gemini and Codex BTW - see how-to). Since working with this my efficiency went way up. A lot of ideas come from @garrytan & @karpathy, while others are mine from my own experience.
Make sure to personalize it and add/remove what makes sense for you. And please LMK if you have ideas to make it better, always happy to learn more :)
dl it here: github.com/jbarbier/CLAUDE.m…
76
routing is table stakes once you hit real token scale, the question is what signals feed it
Jun 3
As token budgets take on a larger part of operating expenses over time, model routing is the inevitable conclusion. This is also one of the biggest areas of differentiation for the applied AI layer over time.
By understanding the different work patterns in your domain, and having strong evals for that domain, you’ll be able to cost/performance optimize effectively.
We’re still likely at the point where most use-cases will need frontier performance for the foreseeable future; but soon you will be able to peel off individual use-cases and send them to lower cost models once the quality is sufficient for the task.
Enterprises individually trying to figure this out themselves at scale will likely not be possible, so the products that can intelligently route these workflows to the right tier of model will be in a strong position to aggregate more demand.
1
3
93
florida naming Altman personally in a state AG lawsuit is the move — 83 pages tying ChatGPT to the FSU shooting and teen suicides, billions in potential penalties, first state to go this far
edition.cnn.com/2026/06/01/b…
51
4x less likely to miss its own code flaws — that's the headline for Opus 4.8, not the benchmarks, the dynamic workflows, or the effort controls
models finally being graded on their blind spots
anthropic.com/news/claude-op…
57
odysseus is what happens when someone builds the whole Claude/ChatGPT UI themselves — chat, agents, email, calendar, deep research, memory, all self-hosted, 7k stars in a few weeks
github.com/pewdiepie-archdae…
73
1 in 5 packages on ClawHub were malicious at peak — agent skills aren't sandboxed like npm deps, they inherit full system permissions, so the blast radius of a typosquatted MCP skill is just different
blog.cyberdesserts.com/ai-ag…
52
base just swapped a 7-day optimistic window for 1-day ZK finality on $7.4B in deposits — SP1 proving the whole thing in Rust, no bespoke circuits, that's the part that matters
blog.succinct.xyz/base-sp1/
39
no cookie, no canvas fingerprint, no webgl — just JavaScript measuring how long your SSD takes to respond while you browse, 89% accuracy on which sites you're visiting, zero permissions required 🤔
tomshardware.com/tech-indust…
32
most params in big AI models are basically zero. why are we spending energy multiplying by them
stanford built Onyx, a chip that skips zeros entirely, 70x less energy and 8x faster than CPUs on average
read more here spectrum.ieee.org/sparse-ai
4
108
this is hilarious
3
187
been poking around stwo - 1400 open branches lol
anyone shipping with it yet? curious how close it is to prod for verifiable ai
github.com/starkware-libs
2
94